An Event-Driven E-Skin System with Dynamic Binary Scanning and real time SNN Classification
Gaishan Li † , Zhengnan Fu † , Anubhab Tripathi, Junyi Yang, Arindam Basu* · 2026
Read An Event-Driven E-Skin System with Dynamic Binary Scanning and real time SNN Classification by Gaishan Li † , Zhengnan Fu † , Anubhab Tripathi, Junyi Yang, Arindam Basu* (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-10537.txt.
Introduction
Gaishan Li and co-authors present an event-driven electronic skin that senses touch and classifies it with a spiking neural network. A piezoresistive array is scanned with an event-based binary search, which slashes scan counts and leaves the data stream almost entirely sparse. That stream feeds a spiking network on an FPGA, using far less compute and storage than a plain CNN. The system still read handwritten digits in real time with high accuracy.
Catalog record
| CREATOR | Gaishan Li † , Zhengnan Fu † , Anubhab Tripathi, Junyi Yang, Arindam Basu* |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 16968 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.10537
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
An Event-Driven E-Skin System with Dynamic Binary Scanning and real time SNN Classification Thanks: This work was supported by the Research Grants Council of the HK SAR, China (Project No. CityU 11212823 ) Gaishan Li † , Zhengnan Fu † , Anubhab Tripathi, Junyi Yang, Arindam Basu* Abstract This paper presents a novel hardware system for high-speed, event-sparse sampling-based electronic skin (e-skin) that integrates sensing and neuromorphic computing. The system is built around a 16×16 piezoresistive tactile array with front-end and introduces a event-based binary scan search strategy to classify the digits. This event-driven strategy achieves a 12.8× reduction in scan counts, a 38.2× data compression rate and a 28.4× equivalent dynamic range, a 99% data sparsity, drastically reducing the data acquisition overhead. The resulting sparse data stream is processed by a multi-layer convolutional spiking neural network (Conv-SNN) implemented on an FPGA, which requires only 65% of the computation and 15.6% of the weight storage relative to a CNN. Despite these significant efficiency gains, the system maintains a high classification accuracy of 92.11% for real-time handwritten digit recognition. Furthermore, a real neuromorphic tactile dataset using Address Event Representation (AER) is constructed. This work demonstrates a fully integrated, event-driven pipeline from analog sensing to neuromorphic classification, offering an efficient solution for robotic perception and human-computer interaction. I Introduction Tactile sense is increasingly pivotal for enabling machines to interact physically and intelligently with their environment, finding critical applications in robotics for dexterous manipulation and in human-computer interaction for rich, physical interfaces [1], [2]. Among the various sensing technologies, resistive-based tactile systems offer significant advantages in terms of scalability and cost-effectiveness, facilitating their deployment over large areas [3]. These systems typically consist of arrays of sensor elements where applied pressure causes a measurable change in resistance, which is then digitized through scanning circuits and an analog front-end (AFE). However, conventional frame-based scanning methodologies—which drive most current systems—suffer from fundamental efficiency bottlenecks [4], [5]. Inspired by imaging sensors, these approaches periodically poll all sensors in the array to construct a full ”tactile frame” [6], yet they are inherently mismatched to the sparse spatiotemporal nature of tactile data. By continuously acquiring data from every sensor regardless of activity, such systems generate substantial redundancy, leading to excessive power consumption and high transmission overhead [7]. Although some studies have employed more complex algorithms to achieve contact timing detection and high-quality tactile image reconstruction, the practical deployment of these solutions in edge electronic skin systems remains challenging [8]. The field of dynamic tactile sensing and e-skin continues to face fundamental challenges beyond data acquisition: while event-driven scanning reduces data at the source, most systems still rely on conventional artificial neural networks (ANN) for classification[9,10]. This creates a critical mismatch: the sparse, event-based data must be reformatted into dense frames for ANN processing, undermining the efficiency gains from dynamic scanning[11]. ANNs further suffer from substantial computational demands and parameter counts that are ill-suited for embedded systems. In contrast, spiking neural networks (SNN) offer a more suitable alternative, maintaining the event-based nature throughout the processing pipeline while achieving higher compression rates and requiring fewer parameters [12]. This end-to-end neuromorphic approach, combining event-driven acquisition with SNN processing, remains largely unexplored in current e-skin implementations. Fig. 1: (a) Proposed overall neuromorphic hardware system, (b) Data heatmap, (c) Event-based Binary Scan Search Strategy (Pulse display of the 16 strongest channels), (d) Convolutional spiking neural network structure used in this system Our work addresses these limitations through a bio-inspired neuromorphic system that co-designs an event-driven scan strategy with efficient SNN processing. As illustrated in Fig. 1(a), the system comprises a sensor array, AFE, FPGA, and PC, forming a complete tactile data acquisition and processing pipeline. Tactile events are usually sparse and short-lived and the heat map of a tactile event is shown in Fig. 1 (b). We develop the event-based binary scan strategy in Fig. 1 (c) efficiently locates these active regions and results sparse, dynamic tactile signals are converted into spikes via delta modulation. The Conv-SNN architecture implemented on the FPGA (Fig. 1 (d)) processes these asynchronous event streams to efficiently extract temporal patterns with low neuronal activity, aligning perfectly with the sparse nature of the tactile data. The principal innovations of this work are following: (1) a novel event-driven binary scan search strategy that increases data sparsity by 7.6×, a 12.8× reduction in scan counts and a 38.4× data compression rate (\Delta=6 ); (2) a Conv-SNN on FPGA which requires only 65% of the computation and 15.6% of the weight storage of a conventional CNN while delivering a 92.11% classification accuracy after quantilization (5bit); and (3) the establishment of a complete event-driven tactile processing pipeline, supported by a real neuromorphic AER dataset, which integrates analog sensing with neuromorphic computing. This work is also delivered to live demonstration track. II Sensor, front-end and scanning algorithm II-A Sensor and front-end hardware system The distributed flexible pressure sensor array is fabricated via precision printing of nanomaterial-based force-sensitive layers and silver interconnects on a thin-film substrate shown in Fig. 1 (a). Configured as a 16×16 array within a 150 mm × 150 mm area, the sensor contains 256 units, each measuring 7.5 mm × 7.5 mm. The sensor exhibits piezoresistive behavior, with resistance decreasing as applied pressure increases. Calibration using a robotic arm confirms that resistance follows an inverse relationship with pressure, varying consistently from several GΩ down to 3 kΩ across the tested range. This well-defined characteristic provides key guidance for front-end amplifier design, ensuring appropriate dynamic range and readout resolution. Fig. 2 illustrates a multi-channel data acquisition system built around an FPGA and a 16×16 resistive sensor array. Each sensor element is integrated into a parallel negative feedback circuit. The system employs switch chips (ADG734) controlled by a register (74HC16D) to connect array rows either to a precision reference voltage V_{ref} or to GND. Selecting a row by connecting it to GND activates a feedback amplifier configuration using OPA2991 operational amplifiers. This switchable architecture enables multiple monitoring modes, enhancing the system’s data compression capability. The output voltage is defined by V_{o}=V_{ref}(1+R_{f}/R_{s}) , where the feedback resistor R_{f} = 100 kΩ and the sensor resistance R_{s} spans from several kΩ to hundreds of GΩ. To ensure stability against parasitic capacitance effects from the high-impedance source, a compensation capacitor C_{f} is implemented across R_{f} for Miller compensation. Fig. 2: Schematic of the readout hardware system of tactile sensors The signal chain utilizes eight low-noise, rail-to-rail OPA2991 op-amps to fully exploit the ADC’s dynamic range. A high-stability REF5025 voltage reference, with a temperature coefficient of 3 ppm/°C, provides the reference V_{ref} . The amplified analog signals are digitized by a 16-channel ADS7961 ADC capable of sampling at 1 MHz, with a programmable input range of 0 to V_{ref} or 0 to V_{ref}/2 . The digitized data is subsequently transmitted to the FPGA via the SPI communication protocol for further processing. II-B Event-based Binary Scan Search Strategy As the integration density of tactile sensors increases, few tactile arrays with over a thousand sensing elements achieve sampling rates exceeding 100 Hz[13]. For the traditional one-by-one scanning method, the average number of scans is \frac{N}{2} which restricts the application of high spatial resolution tactile sensors. However, the spatial sparsity and